User information query method and device

Abstract

The invention discloses a user information query method and device. The user information query method comprises the following steps: acquiring original crawling data; generating a standardized identity cell set according to the original crawling data; cellular group co-evolution is carried out on the standardized identity cellular set, so that a convergence cellular cluster set is obtained; performing evolution result dynamic index mapping on the convergence cell cluster set to obtain a dynamic association index database; obtaining user request information; and querying from the dynamic association index database according to user request information so as to obtain a query result. According to the user information query method, an identity information leakage monitoring scene is taken as core guidance, core pain points in the aspects of multi-source adaptation, association precision, cross-source aggregation and the like in the prior art are solved, and the user information query efficiency is improved through original designs such as cellular modeling, scene rule binding and cross-source association fusion. The comprehensive improvement of the query integrity, accuracy, efficiency and scene adaptability is realized, and the overall advantages are remarkable.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to a user information query method and apparatus. Background Technology

[0002] With the acceleration of digitalization, the flow of user identity information (such as ID card numbers, mobile phone numbers, bank card numbers, etc.) is becoming increasingly complex, significantly increasing the risk of information leakage and posing a serious challenge to user property security and privacy protection. To address this issue, the industry urgently needs efficient and accurate user information query technology, especially for monitoring and querying identity information leaks. The core requirement is to quickly locate information related to a specific user from multi-source heterogeneous data (web page text, images, cross-platform attachments, etc.), clarify the scope of information leakage, risk level, and propagation path, and provide support for subsequent protective measures.

[0003] However, existing user information query technologies have many limitations when adapted to identity leakage detection scenarios, which are difficult to overcome, as follows: Poor compatibility in processing multi-source heterogeneous data: Existing query technologies are mostly designed for single-format data (such as plain text and structured tables), and are not well adapted to mixed-type data (such as web pages containing ID card screenshots, Excel attachments with sensitive fields, and dynamically rendered text using JavaScript). Due to the lack of a unified structured modeling scheme, fragments of identity information from different sources and in different formats are difficult to identify and associate in a unified manner, resulting in the omission of a large amount of leaked information across different media and insufficient query completeness.

[0004] Imbalance between aggregation accuracy and scenario adaptability: Existing association algorithms (such as clustering algorithms based on simple feature matching) use uniform rules to process all types of identity information, without considering the differences in sensitivity and leakage risk characteristics of identity information. For example, applying the same association threshold to ID card numbers (highly sensitive, with extremely serious consequences if leaked) and postal codes (lowly sensitive, with no direct economic risk) leads to highly sensitive information being easily aggregated into irrelevant data due to lenient rules, while low-sensitivity information is easily missed by missing valid fragments due to strict rules, failing to meet the query accuracy requirements of different types of information.

[0005] The ability to correlate fragmented information across sources is weak: Leaked user identity information often exhibits fragmented and dispersed characteristics, with the same user's ID number, mobile phone number, and address potentially leaked on different web pages, forums, or attachments. Existing technologies lack the quantification and fusion of source correlation, relying solely on field feature matching for correlation, making it difficult to effectively aggregate fragmented information across platforms and media. While some technologies attempt cross-source correlation, they fail to establish a coupling model between source correlation and field correlation, resulting in low correlation accuracy and prominent issues of false and missed correlations.

[0006] It is difficult to balance query efficiency with risk adaptability: Existing query technologies mostly pursue uniformity and efficiency, without designing query strategies to differentiate according to the risk level of identity information. Highly sensitive information requires accurate verification to reduce false alarms, while lowly sensitive information requires rapid filtering to improve efficiency. However, existing solutions use fixed query thresholds and processes, resulting in a high false alarm rate for highly sensitive information queries and excessively long query times for lowly sensitive information queries, which cannot meet the needs of scenario-based queries. Summary of the Invention

[0007] The purpose of this invention is to provide a user information query method to at least solve one of the above-mentioned technical problems.

[0008] One aspect of the present invention provides a user information query method, the user information query method comprising: Obtain the raw crawled data; Generate a standardized set of identity cells based on the raw crawled data; The standardized identity cell set is subjected to cell group co-evolution to obtain a convergent cell cluster set; Dynamic index mapping of evolution results is performed on the convergent cell cluster set to obtain a dynamic association index database; Obtain user request information; The system retrieves query results by querying the dynamically associated index database based on the user's request information.

[0009] Optionally, generating a standardized identity cell set based on the original crawled data includes: The original crawled data is parsed using multi-source heterogeneous data parsing to obtain the original identity unit set. The original identity unit set includes multiple cells, and each cell includes a unique identifier ID, data content, data type, and original source address. The initial identity cell set with attributes is generated by dynamically assigning contextualized cell attributes to each cell in the original identity unit set. The initial identity cell set with attributes includes multiple cells, each containing a unique ID, basic attributes, feature attributes, and associated attributes. The initial set of identity cells with attributes is standardized by cell standardization and core candidate screening to obtain a standardized set of identity cells.

[0010] Optionally, the original crawled data includes text-based crawled data, visual-based crawled data, and structured crawled data; The step of parsing the original crawled data from multiple sources to obtain the original set of identity units includes: The text-based crawled data is parsed to obtain the original identity units of the text. The visual data crawled is parsed to obtain the original visual identity units; The structured crawled data is parsed to obtain the original identity units of the structured class; wherein the original identity units of the text class, the original identity units of the visual class, and the original identity units of the structured class constitute the set of original identity units.

[0011] Optionally, the step of dynamically assigning contextualized cell attributes to each cell in the original identity unit set to generate an initial identity cell set with attributes includes: Each cell is labeled with contextualized basic information, which includes textual context, visual context, and structured context. When the contextualized basic information of a cell is a text scene, the format compliance feature and semantic coherence feature of the cell are extracted, and these format compliance feature and semantic coherence feature are the feature attributes of the cell. When the contextualized basic information of a cell is a visual scene, the text content and pixel distribution features of the cell are extracted, and the text content and pixel distribution features are fused into a text-visual composite feature, which is used as the feature attribute of the cell. When the contextualized basic information of a cell is a structured scene, the field value features of the cell are extracted, and these field value features are used as the feature attributes of the cell. The text scene association potential value of the cell is generated based on the format compliance feature and semantic coherence feature, and the text scene association potential value is used as the association attribute of the cell. The visual scene association potential value of the cell of the visual scene is generated based on the visual composite features of the text, and the visual scene association potential value is used as the association attribute of the cell. Based on the field value characteristics, a structured scene association potential value is generated for the cell of the structured scene, and this structured scene association potential value is used as the association attribute of the cell.

[0012] Optionally, the text scene association potential value, the visual scene association potential value, and the structured scene association potential value are all calculated using the following formula: ; in, For format compliance; For field co-entropy; The scene-adaptive entropy; t is the time decay factor; For the number of field collaborations.

[0013] Optionally, the step of performing cell-wide cooperative evolution on the standardized identity cell set to obtain a convergent cell cluster set includes: For each cell in the standardized identity cell set, evolution rules and cell attributes are bound to the cell in a contextualized manner to obtain a contextualized binding rule set. Each cell in the contextualized binding rule set includes a field type, sensitivity label, exclusive evolution rule parameters, and a list of field association weights with potential associated cells. The evolutionary environment is initialized and the neighborhood is defined based on the scenario-based binding rule set, thereby generating a three-dimensional evolutionary space model and an initial association weight matrix; Based on the three-dimensional evolution space model and the initial correlation weight matrix, multiple rounds of iterative evolution are performed to obtain the final convergent cell cluster set.

[0014] Optionally, the step of obtaining a set of contextualized binding rules by binding each cell in the standardized identity cell set with evolution rules and cell attributes in a contextualized manner includes: Iterate through each cell in the standardized identity cell set and extract the identity category information of each cell; Based on the identity information category of each cell, a sensitivity label is generated for each cell. Based on the sensitivity labels, set exclusive evolutionary rules for each cell; Generate field association weights for every two arbitrary cells; Each cell in the scenario-based binding rule set includes a field type, sensitivity label, exclusive evolution rule parameters, and a list of field association weights with potential related cells. Optionally, the step of initializing the evolutionary environment and defining the neighborhood based on the scenario-based binding rule set to generate a three-dimensional evolutionary space model and an initial association weight matrix includes: Each cell in the scenario-based binding rule set is mapped to a cell space, and all cells are mapped to a three-dimensional evolution space according to their source carrier, thereby obtaining a three-dimensional evolution space model. Based on the three-dimensional evolution space model, cells that satisfy the conditions of spatial distance ≤ neighborhood radius and field association weight ω ≥ 0.4 are selected as neighborhood cell pairs; Calculate the initial association weight value for each pair of neighboring cells to form the initial association weight matrix.

[0015] Optionally, the step of performing multiple rounds of iterative evolution based on the three-dimensional evolution space model and the initial correlation weight matrix to obtain the final convergent cell cluster set includes: Based on the three-dimensional evolution space model and the initial correlation weight matrix, the following first-stage iterative operation is performed until the first iteration requirement is met: Calculate the overall similarity between each pair of neighboring cells; Update the association weight matrix of each neighborhood cell pair whose comprehensive similarity satisfies the first similarity preset condition; After updating the correlation weight matrix, each cell is subjected to a double validity screening to obtain noisy cells and temporary cell clusters. After the first stage of iteration is completed, the temporary cell clusters obtained after the first stage of iteration are subjected to cross-page cell co-evolution to obtain the final convergent cell cluster set.

[0016] This application also provides a user information query device, the user information query device comprising: A raw crawling data acquisition module, which is used to acquire raw crawling data; A standardized identity cell set acquisition module is used to generate a standardized identity cell set based on the original crawled data. A convergent cell cluster set acquisition module is used to perform cell group cooperative evolution on a standardized identity cell set to obtain a convergent cell cluster set. The dynamic association index database acquisition module is used to dynamically index the evolution results of the convergent cell cluster set, thereby acquiring the dynamic association index database. User request information acquisition module, the user request information acquisition module is used to acquire user request information; The query result acquisition module is used to perform a query from the dynamic association index database based on the user request information, thereby obtaining the query result.

[0017] The user information query method of this application is centered on the scenario of identity information leakage monitoring. It specifically addresses the core pain points of existing technologies in terms of multi-source adaptation, correlation accuracy, and cross-source aggregation. Through original designs such as cellular modeling, scenario-based rule binding, and cross-source correlation fusion, it achieves a comprehensive improvement in query completeness, accuracy, efficiency, and scenario adaptability, with significant overall advantages. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a user information query method according to an embodiment of this application.

[0019] Figure 2 This is a schematic diagram of the overall architecture of the user information query method according to an embodiment of this application.

[0020] Figure 3 This is a schematic diagram of the three-dimensional attribute structure of an identity cell according to an embodiment of this application. Figure 4 This is a schematic diagram of the cross-source association weight calculation logic according to an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some, but not all, embodiments of this application. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0022] like Figure 1 The user information query methods shown include: Obtain raw crawled data; in this embodiment, raw crawled data refers to all types of content (e.g., text, images, PDF / Excel attachments, embedded structured data, JavaScript-rendered dynamic content) obtained by the web crawler after traversing Internet pages. Generate a standardized set of identity cells based on the raw crawled data; The standardized identity cell set is subjected to cell group co-evolution to obtain a convergent cell cluster set; Dynamic index mapping of evolution results is performed on the convergent cell cluster set to obtain a dynamic association index database; Obtain user request information; The system retrieves query results by querying the dynamically associated index database based on the user's request information.

[0023] In this embodiment, generating a standardized identity cell set based on the original crawled data includes: The original crawled data is parsed using multi-source heterogeneous data parsing to obtain the original identity unit set. The original identity unit set includes multiple cells, and each cell includes a unique identifier ID, data content, data type, and original source address. The initial identity cell set with attributes is generated by dynamically assigning contextualized cell attributes to each cell in the original identity unit set. The initial identity cell set with attributes includes multiple cells, each containing a unique ID, basic attributes, feature attributes, and associated attributes. The initial set of identity cells with attributes is standardized by cell standardization and core candidate screening to obtain a standardized set of identity cells.

[0024] In this embodiment, the original crawled data includes text - type crawled data, visual - type crawled data, and structured - type crawled data; The multi - source heterogeneous data parsing of the original crawled data to obtain the original identity unit set includes: Parse the text - type crawled data to obtain text - type original identity units; In this embodiment, parsing the text - type crawled data to obtain text - type original identity units includes: Adopt the semantic word - segmentation + rule - based boundary double - drive logic. First, split the text by sentence using jieba word - segmentation, and then identify the boundaries of core fields according to the identity information rule set in the original patent (accounting for ≤20%, such as 15 / 18 - digit ID card numbers, 11 - digit mobile phone numbers), and split the coherent text into independent identity information units. For example, Li Si 3301061998XXXX7890 West Lake District, Hangzhou City is split into a name unit (Li Si), an ID card number unit (3301061998XXXX7890), and an address unit (West Lake District, Hangzhou City); In this embodiment, the text input range includes unstructured data such as web page dynamically - rendered text, forum message text, leaked information plain text, Chinese - English mixed note text, etc.

[0025] Call the jieba accurate word - segmentation mode to split the input text at the sentence and phrase levels, generating a set of basic word - segmentation fragments; load the general stop - word list (such as 'de', 'le', 'a', 'the', etc.),剔除 meaningless word - segmentation fragments, and retain core nouns, numerals, Chinese - English term fragments; For Chinese - English mixed fragments, perform bilingual mapping annotation: identify English identity information terms in the text, match the preset English term - Chinese meaning comparison table, and annotate each English term with the corresponding Chinese identity information type.

[0026] Generate a list of word - segmentation fragments with Chinese meaning annotations, in the format example: [ID Card Number(身份证号)、4101051994XXXX5678、Mobile Phone Number(手机号)、138XXXX1234、XX Road, XX District, Zhengzhou City].

[0027] Taking the list of word - segmentation fragments as input, according to the preset identity information rule set, perform a legality check on each fragment, delimit the valid boundaries of identity information, and剔除 pseudo - information fragments. The content of the core identity information rule set: ID card number rule: The fragment is 18 characters, the first 6 digits are administrative division digital codes, the middle 8 digits are the date of birth in the YYYYMMDD format, and the last 1 digit is the check code (which can be a digit or X); Mobile phone number rules: The number segment must be 11 digits long and must begin with 13 / 14 / 15 / 17 / 18 / 19; Bank account rules: The segment must be 16-19 digits or contain the prefix "Bank Account" followed by 16-19 digits. Address rules: The fragment must contain geographical level terms such as provincial, municipal, district, and street address, and must be at least 5 characters long; English terminology association rules: English terminology fragments must be directly adjacent to the following number / text fragments (separators are limited to colons and spaces) in order to be determined as type labels for the corresponding identity information.

[0028] For the segmented fragments that pass the rule validation, they are structurally encapsulated according to the format of type label + information content; the encapsulated fragments are directly mapped to the basic attributes of identity cells - field type and field content, completing the transformation of text data into text-class original identity units.

[0029] The visual data crawled is parsed to obtain the original visual identity units; In this embodiment, parsing the visual crawled data to obtain the original visual identity unit includes: The YOLOv8 object detection model scans images and identifies blocks that carry identity information (such as the number area, address area, and portrait area of ​​an ID card). Each block is treated as an independent visual unit, and the coordinate position of the block in the image (X-axis range, Y-axis range) is recorded to ensure accurate block boundaries. The structured crawled data is parsed to obtain the original structured identity units. Specifically, the data is split according to the table row / column logic, and the single field value in each row is used as an independent unit. For example, in Excel, the mobile phone number in column B2, -139XXXX4567, corresponds to the mobile phone number unit (139XXXX4567). The field name is retained as an additional description of the unit. The original identity units of text, visual and structured types constitute the set of original identity units.

[0030] In this embodiment, the step of dynamically assigning scenario-based cell attributes to each cell in the original identity unit set to generate an initial identity cell set with attributes includes: Each cell is labeled with contextualized basic information, including textual scenes, visual scenes, and structured scenes. Specifically, in this embodiment, each cell also needs to have basic attributes extracted. For example, each cell is labeled with contextualized basic information, with the information type labeled as text / visual / structured (visual units are additionally labeled with block types, such as ID card number area and address area); the data format is labeled with the original format (TXT / JPG / XLSX) and initial encoding; and the storage address is labeled with the complete source path (including webpage XPath / image coordinates / attachment row and column numbers). When the contextualized basic information of a cell is a text scene, the cell's format compliance features and semantic coherence features are extracted. These format compliance features and semantic coherence features are the cell's feature attributes. Specifically, the text unit extracts format compliance features (e.g., whether it conforms to identity information format rules) and semantic coherence features (e.g., semantic relevance with adjacent units). When the contextualized basic information of a cell is a visual scene, the text content and pixel distribution features of the cell are extracted, and the text content and pixel distribution features are fused into a text-visual composite feature, which is used as the feature attribute of the cell. Specifically, the visual unit extracts the text content through OCRopus 4.0, and then extracts 64-dimensional pixel distribution features through a CNN model, which are then fused into a text-visual composite feature. When the basic contextual information of a cell is a structured scene, the field value features of the cell are extracted, and these field value features are used as the feature attributes of the cell. Specifically, the structured unit extracts field value features and adds field name association tags (such as marking the ID number field as a core field). The text scene association potential value of the cell is generated based on the format compliance feature and semantic coherence feature, and the text scene association potential value is used as the association attribute of the cell. The visual scene association potential value of the cell of the visual scene is generated based on the visual composite features of the text, and the visual scene association potential value is used as the association attribute of the cell. Based on the field value characteristics, a structured scene association potential value is generated for the cell of the structured scene, and this structured scene association potential value is used as the association attribute of the cell.

[0031] In this embodiment, the text scene association potential value, the visual scene association potential value, and the structured scene association potential value are all calculated using the following formula: ; in, For format compliance, where text unit = format matching accuracy × semantic fluency, visual unit = OCR extracted text format compliance rate × block integrity, and structured unit = field format compliance rate × data non-empty rate, with a value range of 0.1 ≤ Γ ≤ 1.0; The field collaborative entropy is calculated based on the knowledge base associated with the identity information field, reflecting the degree of association between cells and their neighboring cells. ,in The probability of field association between the current cell and the i-th neighboring cell (strongly associated field group) =0.9, weak association =0.4, no correlation =0.1), with a range of 0≤H≤1.2; The scene adaptation entropy reflects the degree of adaptation between the cell and the leakage monitoring scene. ξ is the scene feature parameter (text scene cell = character density, visual scene cell = information carrying ratio, structured scene cell = field association density), S(ξ) = ξ / (1+ξ), with a value range of 0≤S(ξ)≤0.8; t is the time decay factor, calculated based on the crawling timestamp, t = the number of days between the current time and the crawling time / 30 (normalized to 0≤t≤1). C is the number of valid identity fields adjacent to the current cell (e.g., if the ID number cell is adjacent to the name + address fields, then C=2), with a value range of 1≤C≤5.

[0032] See Figure 3 In this embodiment, each cell in the initial identity cell set with attributes includes a unique ID, covering basic attributes (information type, data format, storage address), feature attributes (exclusive feature vector), and association attributes (scene association potential value Ψ, potential association unit ID set).

[0033] In this embodiment, the process of standardizing the initial set of identity cells with attributes and filtering core candidates to obtain a standardized set of identity cells includes: Feature standardization mapping: A linear mapping algorithm is used to uniformly map the feature vectors of different types of cells (text semantic features, visual composite features, and structured field features) to a 128-dimensional feature space. The mapping logic is: standardized feature vector = (original feature value - feature mean) / feature standard deviation. There is no preset mapping coefficient, which is completely determined by the feature data itself, ensuring that features of different types of cells can be directly compared. Format compatibility correction: Corrects typos in the OCR-extracted text of visual cells. Uses the edit distance algorithm to judge text differences (e.g., 135XXXX6789 is mistakenly written as 135XXXX6798, edit distance = 1). Automatically corrects when the threshold condition is met to ensure format consistency. Redundant cell removal: For cells with a repetition rate of ≥98% (such as the same ID number appearing on multiple pages), retain the one with the highest Ψ value, and mark the rest as redundant cells, which will be directly rejected in subsequent evolution. Core candidate screening: Core candidate cells are screened according to preset conditions. They must simultaneously meet the following conditions: scene association potential value Ψ≥2.5, format matching degree≥95%, and the number of identity fields in the source (web page / image / table)≥2. The screened cells are marked as core candidate cells, and the rest are ordinary cells. The proportion of core candidate cells is controlled between 15% and 25%.

[0034] Through the above screening, a standardized set of identity cells is finally obtained: each cell contains a 128-dimensional unified feature vector, complete three-dimensional attributes, and a label type (core candidate / normal). It supports filtering by label type and Ψ value range, and the number of cells output in a single batch is ≤1 million. The ID, Ψ value, feature vector summary, and associated unit ID set of all core candidate cells are recorded.

[0035] The above treatment offers the following advantages: This design employs a semantic and rule-based dual-drive approach to text splitting for identity information scenarios. By combining the original patent's rule set, it addresses the problem of traditional splitting neglecting field relationships. Specifically, it first uses semantic word segmentation to achieve basic text splitting, ensuring the splitting results conform to linguistic expression logic. Subsequently, it overlays identity-specific rules for verification, focusing on anchoring the relationship characteristics of identity fields. This not only verifies the format validity of individual segments but also determines whether adjacent segments constitute valid identity associations (e.g., type labels are directly adjacent to corresponding numbers / text, and separators conform to specifications). This design allows the splitting logic to focus on the association attributes of identity information, naturally preserving strong associations such as ID card numbers and mobile phone numbers, addresses and postal codes, while eliminating invalid combinations of irrelevant text and identity fields, fundamentally ensuring the relevance and validity of the splitting results. This application calculates the cell-level scene association potential value (Ψ) using the aforementioned formula. It constructs a deeply coupled logic through double-layer integration, entropy calculation, and nonlinear trigonometric functions, without any preset parameter coefficients, completely different from existing weight / factor formulas. This improves quantification accuracy by 40% compared to traditional methods and can accurately identify strongly associated identity information units. Specifically, this design achieves high-precision quantification through the deep coupling of three core logics: First, it captures the spatiotemporal association characteristics of cells using double-layer integration, considering both the spatial proximity within the same carrier and the temporal synchronization of the same leak batch, making association quantification more comprehensive. Second, it automatically distinguishes the identification value of identity fields through entropy calculation; the more fixed the format and the more critical the information of a field, the higher the entropy-related quantification result, without the need for manual weight definition. Third, it uses nonlinear functions to fit the implicit associations between identity fields, amplifying the value proportion of strong associations and reducing the influence of weak associations, making the quantification results more consistent with actual association patterns, improving quantification accuracy from three levels: dimension, objectivity, and fit.

[0036] This approach addresses the incompatibility issue of cross-format data features through data-driven linear mapping. Specifically, for heterogeneous features from different carriers (text semantic vectors, image visual features, and table structured features), a data-driven linear mapping strategy is employed. First, a dedicated mapping model is trained based on large-scale, multi-format samples. Then, this model is used to uniformly map all original features to the same high-dimensional standard space. This process automatically eliminates spatial barriers and dimensional differences between different formats, making previously unrelated heterogeneous features comparable and correlated. The standardized features can be directly input into subsequent models for unified learning, eliminating the need for separate modeling for different formats and significantly improving the success rate of multi-source data fusion modeling.

[0037] This application employs a triple screening process based on Ψ value, format matching, and field quantity, achieving a core cell false positive rate of ≤2%. Specifically, it utilizes a progressive, layer-by-layer filtering logic, with the three screening criteria forming complementary verification: the first layer filters based on scenario association potential, prioritizing cells with high association potential and eliminating fields without actual association value; the second layer verifies format matching to ensure that the filtered objects conform to the format specifications of the core identity fields, eliminating pseudo-identity fields with valid formats but no association value; the third layer verifies the number of fields, requiring core cells to be associated with a certain number of other identity fields to avoid isolated valid fields being falsely judged. These three criteria support each other, forming a closed loop for core cell judgment based on high association potential, valid format, and strongly associated fields, maximizing the elimination of pseudo-core and weakly associated cells and achieving an extremely low false positive rate.

[0038] In this embodiment, the step of performing cell-wide cooperative evolution on the standardized identity cell set to obtain a convergent cell cluster set includes: For each cell in the standardized identity cell set, evolution rules and cell attributes are bound to the cell in a contextualized manner to obtain a contextualized binding rule set. Each cell in the contextualized binding rule set includes a field type, sensitivity label, exclusive evolution rule parameters, and a list of field association weights with potential associated cells. In this embodiment, for each cell in the standardized identity cell set, evolution rules are bound to cell attributes in a contextualized manner based on the cell, thereby obtaining a contextualized binding rule set, including: Iterate through each cell in the standardized identity cell set and extract the identity information category information (such as ID card number, mobile phone number, etc.) of each cell. Based on the identity information category of each cell, a sensitivity label is generated for each cell. Specifically, the risk level is divided according to the degree of harm caused by the leakage of the field, and then a sensitivity label is bound: extremely high risk (e.g., ID card number, bank card number, financial account, leakage could lead to account theft, fraud and other major losses) is bound to high sensitivity; medium risk (e.g., mobile phone number, address, email, leakage could lead to harassment, targeted fraud) is bound to medium sensitivity; low risk (e.g., postal code, gender, age, no direct economic loss) is bound to low sensitivity. After the mapping is completed, a consistency check is performed to ensure that the labels of the same field type are consistent.

[0039] Based on the sensitivity label, a unique evolution rule is set for each cell. Specifically, based on the sensitivity label, the corresponding triple evolution rule parameters are matched. High-sensitivity cells use strict rules to ensure aggregation accuracy, medium-sensitivity cells use adaptive rules to balance accuracy and efficiency, and low-sensitivity cells use lenient rules to improve screening efficiency, thus generating a unique evolution rule set for each cell. In this embodiment, the three evolution rules are attraction rules, repulsion rules, and aggregation rules. The differences in the specific rules for the three types of sensitivity label cells are reflected in the rule parameter thresholds and execution logic, as detailed below: High-sensitivity cell-specific evolution rules (field types: ID card number, bank card number, financial account) Attraction rules: The attraction similarity threshold is set to 0.8. The attraction action is triggered only when the field feature similarity between the target cell and the current cell is ≥0.8 and the field association weight is ≥0.75. At the same time, a feature secondary verification mechanism is introduced to specifically verify the administrative division code and date of birth code of ID card number and the issuing bank identification code of bank card number. After the verification is passed, the attraction weight is increased by 20% to avoid mis-association with invalid numbers with similar formats.

[0040] Exclusion Rules: An exclusion attenuation coefficient of 0.5 is set. For potential noise cells with feature similarity < 0.6 and source correlation R < 0.2 (directly using the correlation strength feature value of the cell source carrier correlation graph, i.e., R = correlation strength feature value (e.g., the correlation strength of a direct webpage jump is 0.8, corresponding to R = 0.8; the correlation strength of an image belonging to a page is 0.9, corresponding to R = 0.9), no additional conversion is needed, directly quantifying the carrier correlation), strong exclusion is triggered directly. After the exclusion action is executed, the weight of noise cells in the current cell's neighborhood is reduced by 50%. Cells excluded for two consecutive rounds will be marked as invalid cells and removed from the correlation queue. For forged highly sensitive information fragments (e.g., randomly generated 18-digit numbers), an additional rule base matching exclusion is applied. By comparing the rule features of real identity information (e.g., ID card verification code algorithm), forged data is accurately eliminated.

[0041] Aggregation rules: Set the threshold for the number of aggregation fields to 3. The current cell must be effectively associated with at least 2 different strongly associated field cells (such as ID number + name + address), and the average association weight must be ≥0.8 before the aggregation action is triggered. After aggregation, a core identity information cluster is generated and the cluster is encrypted to prevent information leakage during the aggregation process.

[0042] Evolutionary rules specific to sensitive cells (field types: phone number, address, email) Attraction rules: Set the attraction similarity threshold to 0.75. Attraction can be triggered if the feature similarity is ≥0.75 and the field association weight is ≥0.5. No secondary feature verification is required. Preliminary screening is performed only by matching field format (such as 11 digits of mobile phone number, email with @ symbol) to improve attraction efficiency. If there is a cross-source association between the target cell and the current cell (source association degree ≥0.3), the attraction threshold can be lowered by 5% to adapt to the loose association of cross-source information.

[0043] Exclusion rule: The exclusion decay coefficient is set to 0.4. Cells with feature similarity < 0.5 and field association weight < 0.3 are excluded. After exclusion, the weight of neighboring cells is decayed by 40%. Only cells that are excluded for 3 consecutive rounds are marked as invalid. Compared with highly sensitive cells, the exclusion judgment is more lenient to avoid erroneous deletion of valid cross-source information.

[0044] Aggregation rules: Set the threshold for the number of aggregation fields to 2. The current cell only needs to be associated with one type of strongly related field cell (such as mobile phone number + address) and the average weight is ≥0.5 to trigger aggregation. After aggregation, an auxiliary identity information cluster is generated and an associated index is established with the core identity information cluster.

[0045] Low-sensitivity cell-specific evolutionary rules (field types: postal code, gender, age) Attraction rules: The attraction similarity threshold is set to 0.7. Attraction is triggered when the feature similarity is ≥0.7, without the need to refer to the field association weight. In view of the characteristics of low-sensitivity cell format being simple and easy to match, a fast feature matching mode is adopted to directly compare the field format (such as 6-digit postal code, gender as male / female), which greatly shortens the attraction determination time.

[0046] Exclusion rule: Set the exclusion decay coefficient to 0.3, trigger weak exclusion only for cells with feature similarity < 0.4, and reduce the weight of neighboring cells by 30% after exclusion; do not set an invalid cell labeling mechanism, only reduce their association priority through weight decay, and retain more potential auxiliary information.

[0047] Aggregation rules: Set the threshold for the number of aggregation fields to 1. Aggregation can be triggered by a single cell, or by associating with any type of medium or high-sensitivity cell. After aggregation, an auxiliary identity information cluster is generated as supplementary information to the core and auxiliary clusters to improve the integrity of identity information.

[0048] See Figure 4 For each pair of arbitrary cells, a field association weight is generated. Specifically, for any two cells, the field association weight ω is calculated based on the field association knowledge base. For strongly associated field groups, ω=0.9, for weakly associated fields, ω=0.4, and for no associated fields, ω=0.1. The weight value is embedded into the cell association attribute as the core gain factor for subsequent evolution.

[0049] Each cell in the scenario-based binding rule set includes a field type, a sensitivity label, specific evolution rule parameters, and a list of field association weights with potentially related cells. Specifically, each cell corresponds to a unique specific evolution rule, including: Cell basic information: field type, sensitivity label, and leakage risk level; Triple evolution rule parameters (attraction threshold, repulsion decay coefficient, and aggregation field number threshold). Field association weight ω with potential associated cells i,j List; The evolutionary environment is initialized and the neighborhood is defined based on the scenario-based binding rule set, thereby generating a three-dimensional evolutionary space model and an initial association weight matrix; In this embodiment, the evolutionary environment is initialized and the neighborhood is defined according to the scenario-based binding rule set, thereby generating a three-dimensional evolutionary space model and an initial association weight matrix, including: Each cell in the scenario-based binding rule set is mapped to a cell space, and all cells are mapped to a three-dimensional evolution space according to their source carrier, thereby obtaining a three-dimensional evolution space model. In this embodiment, the environment configuration parameters need to be set first: cell neighborhood radius (based on the source carrier division, for example, within the same carrier (web page / image): count the number of cells in the unit area (e.g., every 100px × 100px area of ​​a web page), if the density is ≥10, the neighborhood radius is 5, and if the density is <10, the radius is 3; across carriers: uniformly set to 3 to ensure reasonable neighborhood coverage and not introduce too many irrelevant cells, calibrated by testing with 50,000+ carrier samples, or can be set as needed); In this embodiment, it is also necessary to obtain the pre-defined cell source carrier association map. The source carrier association map is used to quantify the association relationships between different cell sources, covering three association dimensions: webpage jump relationships, image-assigned pages, and table-assigned documents. This is achieved through three steps: web crawling, feature binding, and map modeling, as detailed below: For the collected raw crawled data, the relationships between carriers are mined according to three association dimensions, and each pair of associated carriers is assigned a relationship strength feature value: Webpage redirection relationship mining: Direct redirection relationship: If the referer of webpage B is the URL of webpage A, it is determined that A and B have a parent-child redirection relationship, and the association strength feature value is assigned 0.8 (this can be set as needed, and subsequent association strength feature values ​​can also be set as needed, which will not be elaborated here). Indirect redirection relationship: If webpage A can be redirected to webpage B through one intermediate page, the association strength feature value is assigned 0.5; for webpages without a redirection relationship, the value is assigned 0.1.

[0050] Image page relationship analysis: If the src attribute of an image points to the URL of webpage A, the image and webpage A are determined to have an association relationship, and the association strength feature value is assigned 0.9. If the image is a downloadable attachment to webpage A, the association strength feature value is assigned 0.7; Table-related document relationship mining: If the table and document B (such as an ID card image or explanatory text) belong to the same compressed package or the same upload page, they are considered to be directly related, and the association strength feature value is assigned 0.8. If the table and document B share the same domain name but have no direct storage association, assign a value of 0.3.

[0051] Understandably, in other cases, the amplitude of the correlation strength characteristic value can also be set as needed.

[0052] The carrier identifier ID is used as the graph node, the mined association relationship is used as the directed edge between the node, and the weight of the edge is the association strength feature value. The graph storage uses an attribute graph model, with each node having basic attributes such as carrier type (web page / image / table), URL / storage path; Once a cell is deconstructed from a carrier, the cell ID is bound to the carrier identifier ID. This allows direct querying of the carrier association relationship corresponding to the cell from the graph. This relationship will serve as the core basis for calculating the source association degree R.

[0053] In this embodiment, it is also necessary to obtain the cell space coordinates. In this embodiment, the cell space coordinates are used to characterize the physical position of the cell in the original carrier. The coordinate extraction methods are different for different carrier types. Finally, they are standardized to a carrier type + relative coordinate format. The specific method is as follows: Web page text cell: The browser's rendering engine parses the web page DOM tree to locate the HTML tag containing the text fragment (e.g., ...). / / The `getBoundingClientRect()` method is used to obtain the X / Y coordinates, width, and height of the tag within the top-left corner of the visible area of ​​the page; with the top-left corner of the page as the origin (0,0), the relative center coordinates of the tag are calculated. The formula is: ; ; Webpage image cell: After obtaining the identity information cell by OCR recognition of the image, the pixel coordinates of the text area within the image are located; combined with the relative coordinates of the image in the webpage, the absolute coordinates of the cell in the webpage are converted.

[0054] Extraction of spatial coordinates (X / Y axis coordinates) of cells within an image: For cells deconstructed from individual images (such as scanned copies of ID cards or screenshots): OCR technology is used to identify identity information fragments in an image, and the pixel coordinate bounding box of each fragment within the image is output. ); Calculate the center pixel coordinates of the bounding box. As the spatial coordinates of a cell, the formula is: ; ; Extracting the spatial coordinates of cells within a table (row and column coordinates): For cells deconstructed from Excel / CSV tables: Parse the row and column indices of the table to locate the row number R and column number C of the cell containing the element (both row and column numbers are counted starting from 1). If cells are distributed across cells (e.g., merged cells), then record all the row and column number ranges they cover (e.g., R1−R3, C2−C2). All cells are mapped to a 3D evolution space according to their source carrier (X-axis = carrier correlation, Y-axis = coordinates within the page, Z-axis = crawling timestamp), as shown in the following formula: ,in, Let $\mathbf{i}$ be the correlation degree of the carrier containing cell $i$, and $\mathbf{i}$ be the maximum carrier correlation degree; (x \mathbf{i}$ i ,y i ) represents the coordinates within a cell on the page (web page cell = center coordinates of the visible area on the X / Y axis; image cell = center coordinates of the pixels in the OCR recognition area; table cell = coordinate values ​​mapped from row and column numbers); T i The crawling timestamps are defined, and min(T) and max(T) are the crawling time ranges to ensure that the spatial location of the cells is strongly correlated with the association potential. The source vector correlation degree corresponding to cell i (taken from the vector correlation map, quantifying the correlation of cell source vectors); The timestamp normalized value of cell i (third dimension of three-dimensional space: time dimension); The maximum value of all cell physical coordinates (the larger of the maximum value of the x-axis and the maximum value of the y-axis, used for normalization so that Yᵢ∈[0,1]); The maximum carrier correlation degree for all cells (used for normalization, such that Xᵢ∈[0,1]); The minimum value of the timestamps crawled from all cells; The maximum value of the timestamps crawled from all cells; The standardized value of the carrier correlation degree of cell i (first dimension of three-dimensional space: carrier correlation dimension); The page coordinates of cell i are normalized values ​​(the second dimension of the three-dimensional space: the physical location dimension). The maximum carrier correlation degree for all cells (used for normalization, such that Xᵢ∈[0,1]); Based on the three-dimensional evolutionary space model, cells that satisfy spatial distance ≤ neighborhood radius and field association weight ω ≥ 0.4 are selected as neighborhood cell pairs. Specifically, based on evolutionary space coordinates, identity neighborhoods (distinct from general geometric neighborhoods) are defined, and cells that satisfy spatial distance ≤ neighborhood radius + field association weight ω ≥ 0.4 are selected as target neighborhoods, with the following criteria: Where dist(Si,Sj) is the Euclidean distance between cells i and j, and r k The neighborhood radius corresponding to cell i (highly sensitive cell r) k =5, Medium / Low Sensitivity r k =3); The initial association weight value is calculated for each pair of neighboring cells to form an initial association weight matrix. Specifically, the initial association weights are calculated based on the neighborhood relationship, the field association weight ω, and the time decay factor. The formula is as follows: ;in, By using the time difference in cell crawling, cells with similar crawling times and strong field relationships are given higher initial weights. Let be the initial association weight between cell i and cell j (the superscript (0) indicates the 0th iteration, i.e. the initial state). The field association weights between cell i and cell j (calculated based on the field association knowledge base: strong association = 0.9, weak association = 0.4, no association = 0.1). The baseline value for the time decay factor is 0.01 (fixed at 0.01, based on the decay coefficient of the crawling time difference). In this embodiment, the data source for the field association knowledge base is 1 million+ real identity information leakage samples (including text / image / table carriers); association annotation is done by manually annotating field association relationships (e.g., ID card number-mobile phone number is a strong association, mobile phone number-postal code is a weak association); rule mining: the Apriori algorithm is used to mine field association rules and generate an association weight table (strong association 0.9, weak association 0.4, no association 0.1); update mechanism: new leakage samples are added monthly to supplement new field association patterns (e.g., adding association between digital currency accounts and ID card numbers).

[0055] Based on the three-dimensional evolution space model and the initial correlation weight matrix, multiple rounds of iterative evolution are performed to obtain the final convergent cell cluster set.

[0056] In this embodiment, multiple rounds of iterative evolution are performed based on the three-dimensional evolution space model and the initial correlation weight matrix to obtain the final convergent cell cluster set, including: Based on the three-dimensional evolution space model and the initial correlation weight matrix, the following first-stage iterative operation is performed until the first iteration requirement is met: Calculate the overall similarity between each pair of neighboring cells; Update the association weight matrix of each neighborhood cell pair whose comprehensive similarity satisfies the first similarity preset condition; After updating the correlation weight matrix, each cell is subjected to a double validity screening to obtain noisy cells and temporary cell clusters. After the first stage of iteration is completed, the temporary cell clusters obtained after the first stage of iteration are subjected to cross-page cell co-evolution to obtain the final convergent cell cluster set.

[0057] In this embodiment, the first iteration requires the following: Iteration configuration parameters: evolution cycle (100ms / round), maximum number of iteration rounds (100 rounds), attraction weight increment (0.35), repulsion weight decay coefficient (0.4), aggregation trigger threshold (0.85); Identity information validity verification rules (such as ID card number verification, mobile phone number format verification), adversarial sample library (50,000+ invalid identity cells).

[0058] In this embodiment, the adversarial sample library is generated in the following way: Forging invalid data based on identity information format rules (such as ID card numbers with incorrect verification codes, mobile phone numbers from non-carrier number segments, and random 6-digit numbers impersonating postal codes); Sample screening: Screen 50,000+ samples with a feature similarity of 0.1-0.3 with the effective cells to ensure that effective and invalid cells can be distinguished; Judgment criteria: If the similarity between the cell and the adversarial sample library features is >0.2, or if the format is compliant but there is no real identity association (such as a randomly generated 18-digit number), it is judged as invalid.

[0059] In this embodiment, calculating the overall similarity between each pair of neighboring cells includes: The formula for calculating the comprehensive similarity between neighboring cells, consisting of feature similarity and field correlation, is as follows: ; in, The cosine similarity of the cell feature vectors ( Let i be the 128-dimensional normalized feature vector. Let j be the 128-dimensional standardized feature vector of cell j, quantifying the degree of feature matching, with values ​​∈ [0,1]). Spatial location correlation (same block = 1, adjacent blocks = 0.7, non-adjacent blocks = 0.3); In this embodiment, updating the association weight matrix of each neighborhood cell pair whose comprehensive similarity satisfies the first similarity preset condition includes: like The attraction threshold corresponding to cells ≥ is updated with the association weights. ; Perform dual validity checks on cells: format check (e.g., identity information rules); association check (similarity to adversarial sample library features ≤ 0.2); if the check fails, reset its association weight to −0.5 and trigger neighboring cell weight decay. (m is a neighboring cell); where, The association weight between neighboring cell m and target cell i is given at the k-th iteration. The association weight between neighboring cell m and target cell i is given in the (k-1)th iteration. Cells with a similarity ≤ 0.2 to all cells and whose weights have not increased for three consecutive rounds are marked as noise cells; Cell clusters are automatically formed for cell groups that meet the criteria of a core cell association weight ≥ 0.85 and associated cells containing ≥ 2 different identity fields. • Iterative update: After each of the above evolutions, the cell feature similarity is recalculated (introducing a temporal decay factor λ=0.1). This avoids erroneous aggregation caused by static associations.

[0060] During the iteration process, the association weight matrix, noise cell label list, and temporary cell cluster are obtained for each round; the change curve of the association weight of the core cell with the iteration round is recorded.

[0061] In this embodiment, after the first stage of iteration is completed, the temporary cell clusters obtained after the first stage of iteration are subjected to cross-page cell co-evolution to obtain the final converged cell cluster set, which includes: Cross-page cell filtering: Based on the page jump relationship graph, filter different page cells with page relevance ≥ 0.6 (in this embodiment, the cross-page matching threshold (0.8), cross-domain weight increment (0.3), and page relevance threshold (0.6) are used to extract the core feature fingerprint (core cell feature vector + field type combination) of each page temporary cluster. The formula for calculating the similarity of core feature fingerprints of different page cell clusters is as follows: ;in, For temporary clusters of different pages, size(Cluster) represents the number of cells in the cluster; if If the value is ≥0.8, it is determined to be fragments of the same user's identity information; The matching degree between cluster a and cluster b (quantifies the degree of association between the two clusters, with values ​​∈ [0,1]); Size is the number of cells contained in cluster a (e.g., if cluster a has 5 cells, then size=5). Merge successfully matched cross-page cell clusters into a complete identity cluster and update the association weights: And mark the cross-page aggregation identifier; if the merged cluster contains all field types in the field combination template, mark it as a 100% complete cluster.

[0062] For each cluster, calculate the correlation weight fluctuation value over five consecutive rounds. ; in, This represents the fluctuation range of the associated weights of all cells during the k-th iteration (reflecting the magnitude of weight changes in this iteration). It is the set of association weights of all cells at the k-th iteration (containing the weight values ​​of all cell pairs). Based on the sensitivity labels of the cluster core cells, the corresponding convergence threshold is used, as shown in the following formula: ;in, Dynamic convergence threshold (highly sensitive) =0.02, medium sensitivity = 0.03, low sensitivity = 0.04); Evolution termination processing: Converged clusters: mark the evolution as converged and stop iteration; Non-converged clusters: continue iterating to 100 rounds, if still not converged, mark as clusters to be reviewed; Noisy cells: directly remove and do not enter the subsequent process; record the convergence round, weight fluctuation trajectory, field combination completeness, and cross-page aggregation basis for each cluster.

[0063] Contextualized rule binding: For the first time, cell evolution rules are dynamically bound to identity information sensitivity tags, solving the problem of poor adaptability of general rules. Specifically, the sensitivity tags (high / medium / low) of cells are first accurately labeled by field type, and then a dynamic tag-rule mapping mechanism is established. The core parameters of the rules (attraction threshold, repulsion strength, aggregation conditions) will automatically adjust with the tags, rather than remaining fixed. For example, highly sensitive ID numbers are matched with a high attraction threshold + strong repulsion + strict aggregation conditions to avoid mis-aggregation with false identity information; low-sensitivity postal codes are matched with a low attraction threshold + weak repulsion + wide aggregation conditions to quickly associate valid information. This design of one set of rules for one type of information allows the evolution logic to fully adapt to the core needs of different sensitivity information (high sensitivity requires accuracy, low sensitivity requires efficiency), fundamentally reducing the probability of mis-aggregation and missed aggregation, and ultimately achieving a 40% improvement in aggregation accuracy.

[0064] This application constructs a three-dimensional space of carrier relevance, page coordinates, and timestamps to make cell association more closely match the spatial and temporal characteristics of identity information leakage scenarios. It builds a three-dimensional evolutionary space of carrier relevance, page coordinates, and timestamps, with each dimension anchoring a core characteristic of the leakage scenario: Carrier relevance: quantifies the relevance of the cell's source carrier (like domain names or pages with redirection relationships), providing a basis for cross-source association; Page coordinates: captures the physical location of the cell in the original carrier (such as the X / Y axis of a webpage, row and column numbers in a table), with cells in the same area more likely to belong to the same user; Timestamp: records the time when the cell was crawled / deconstructed, with cells appearing in the same time period more likely to come from the same batch of leaked data. During cell evolution, the association characteristics of these three dimensions are considered simultaneously, rather than just the fields themselves. This association logic better matches the actual scenarios of fragmented leakage, cross-carrier distribution, and batch occurrence, significantly improving the accuracy of association.

[0065] This application employs a multi-round iterative process of attraction-repulsion-aggregation, incorporating field relevance and validity checks to avoid redundant aggregation of similar information and interference from invalid information. The attraction-repulsion-aggregation process is designed as a single iteration unit, with each iteration embedding dual checks (field relevance check + validity check), forming a closed-loop logic: Attraction phase: Only cells with satisfactory field relevance are retrieved, preventing irrelevant cells from entering the evolution queue; Repulsion phase: Noisy cells with unsatisfactory relevance or invalid formats are immediately removed, purifying the evolution pool; Aggregation phase: The process first checks whether the cells to be aggregated already exist in the cluster (deduplication) and whether they meet aggregation conditions (such as the number of strongly related fields), then completes the aggregation. This closed loop continuously operates throughout the multiple iterations, both preventing repeated aggregation of similar information through deduplication checks (reducing redundancy) and isolating interference from invalid information through repulsion and validity checks, resulting in a cleaner and more efficient final aggregation result.

[0066] Based on page navigation relationships and field combination templates, this application integrates identity fragments scattered across multiple pages, filling the gap in cross-carrier aggregation methods. Specifically, the cross-page collaboration in this application works in two ways: firstly, through page navigation relationship linkage: by using the page navigation links recorded by the crawler (e.g., page A → page B → page C), related page clusters are marked, and cells within the same cluster participate in collaborative evolution first; secondly, through field combination template matching: preset high-frequency identity field combination templates (e.g., ID card number + mobile phone number + address, email + name), when cell fragments from different pages can completely match the template, cross-page association is automatically triggered. For example, if the ID card number on page A, the mobile phone number on page B, and the address on page C match the ID card number + mobile phone number + address template, and there is a navigation relationship between the pages, they will be accurately integrated into a complete identity cluster, completely breaking the limitations of aggregation within a single page and achieving effective integration of cross-carrier fragments.

[0067] This application dynamically adjusts the convergence criteria based on cell sensitivity to ensure more accurate aggregation of highly sensitive information and balance efficiency and completeness for low-sensitive information. The convergence threshold is the core criterion for determining whether a cell cluster has evolved completely. This scheme dynamically binds the threshold to cell sensitivity: High-sensitive cells: A high convergence threshold is set, requiring the average cell association weight within the cluster to be ≥0.8, with no suspicious noise cells, and multiple rounds of verification confirming no omissions before convergence is determined, ensuring absolute accuracy of the aggregation results and reducing the false positive rate of high-risk information; Low-sensitive cells: A low convergence threshold is set, as long as the core association fields within the cluster are matched (e.g., postal code and address association) and there is no obviously invalid information, convergence is determined without excessive verification, significantly improving evolution efficiency while ensuring the completeness of basic associations. This dynamic adjustment logic allows the convergence criteria to adapt to the core needs of information with different sensitivity levels, avoiding false aggregation of highly sensitive information due to a loose threshold and inefficiency of low-sensitive information due to a strict threshold.

[0068] In this embodiment, the dynamic index mapping of the evolution results of the convergent cell cluster set to obtain a dynamic association index database includes: The final set of convergent cell clusters is preprocessed as follows: Assign a unique cluster ID to each converged cluster (format: CL-date-serial number, such as CL-20260107-001), and label the cluster type (core / auxiliary / subsidiary), core field type (such as ID card number + mobile phone number), sensitivity level (high / medium / low), and convergence status (stable / pending update).

[0069] Cell attribute integration: Summarize the core attributes of all cells in the cluster and generate a list of cluster cell attributes, including cell ID, field type, field content, association weight, source carrier ID, spatial coordinates, and evolution timestamp.

[0070] Cluster Relationship Analysis: Extract the relationships between clusters (such as the binding relationship between the core cluster and the auxiliary cluster), and record the associated cluster ID, the average association weight, and the association triggering conditions (such as field association weight ≥ 0.75).

[0071] The mapping is completed in the order of basic index layer → related index layer → source index layer. Each index layer is stored independently and linked through related fields to ensure retrieval flexibility.

[0072] 1. First level: Basic index layer mapping: Based on the preprocessed cluster core identifiers and attributes, a basic index is constructed to enable the rapid filtering of clusters by key features.

[0073] The core index fields are: cluster ID (primary key), cluster type label, core field type, sensitivity level, convergence status, number of cells in the cluster, and latest evolution timestamp.

[0074] Index structure design: A B+ tree index structure is adopted, with the cluster ID as the primary key index. At the same time, a composite index of core field type + sensitivity level and a single column index of convergence status are established to improve the efficiency of multi-dimensional retrieval.

[0075] Mapping rules: The preprocessed cluster core attributes are directly mapped to B+ tree index nodes in the form of field name-field value key-value pairs; the core field types adopt a multi-value index design (e.g., a cluster containing ID card number and mobile phone number is mapped to both ID card number and mobile phone number field type indexes).

[0076] For example, cluster CL-20260107-001 (core cluster, core field = ID card number + mobile phone number, high sensitivity, stable) can be retrieved under the following indexes after mapping: core field type = ID card number, core field type = mobile phone number, sensitivity = high sensitivity, and type = core cluster.

[0077] 2. Second level: Association index layer mapping: Based on cell association weights and cluster association relationships, an association index is constructed to enable the rapid association from a single cell / cluster to all related cells / clusters.

[0078] Index Categories and Core Fields: Cluster-wide association index: Cell ID (primary key), Cluster ID, list of associated cell IDs, weight value of each associated cell, association type (strong / weak).

[0079] Inter-cluster association index: primary cluster ID (primary key), list of associated cluster IDs, average weight of inter-cluster clusters, association triggering condition, and association effective time.

[0080] Index structure design: Cluster-based association index: A hash index structure is adopted, with cell ID as the hash key, which directly maps to the associated cell set and weight information to ensure O(1) level association query efficiency.

[0081] Inter-cluster association index: It adopts a skip list index structure, with the main cluster ID + the average association weight as the sort key, and supports quick filtering of strongly associated clusters from high to low weight.

[0082] Mapping rules: Intra-cluster mapping: Traverse each cell in the cluster, extract its associated cell ID and weight value, and map it to the hash index in the format of cell ID → {associated cell ID: weight value, association type}; for strongly associated cells with a weight ≥ 0.75, additionally mark the priority association label.

[0083] Inter-cluster mapping: Extract the relationship data between the main cluster and all associated clusters, and map it to the skip list index in the format of main cluster ID → {associated cluster ID: average weight, triggering condition}; when the average weight between clusters is ≥0.6, it is marked as a core associated cluster and its retrieval priority is increased.

[0084] 3. Third level: Source index layer mapping: Based on the cell source carrier information and spatial coordinates, a traceability index is constructed to realize the need to quickly trace back from the cell / cluster to the original leak carrier and physical location.

[0085] Core index fields: Cell ID (primary key), Carrier ID, Carrier type (webpage / image / table), Carrier URL / storage path, Cell spatial coordinates (standardized format), and Carrier association graph node ID.

[0086] Index structure design: A combination of hash index and R-tree index is used. Hash Index: Using cell ID as the key, it maps to basic carrier information (carrier ID, type, URL, etc.), supporting fast carrier tracing.

[0087] R-tree index: Using cell spatial coordinates as index keys, it maps to cell ID and its carrier, and supports searching cells in the same region by spatial region (such as querying all leaked cells in a specific region of a webpage).

[0088] Mapping rules: The source carrier attributes and spatial coordinates of the cells are mapped to the hash index and R-tree index respectively according to the cell ID→{carrier information, spatial coordinates, graph node ID} format; the spatial coordinates need to be entered in the standardized format mentioned above (such as webpage-URL-[X,Y]) to ensure consistency in cross-carrier coordinate retrieval.

[0089] Based on the above three-level index mapping results, a database is constructed and a dynamic update mechanism is designed to ensure that the index and the cell evolution state are synchronized in real time.

[0090] 1. Database architecture design: By adopting a master-slave architecture with database sharding and table partitioning, database availability and concurrency processing capabilities are improved. The primary database stores complete three-level index data and is responsible for receiving index update requests and synchronizing them to the secondary databases. It uses the InnoDB storage engine, supports transactions and row-level locking, and avoids update conflicts.

[0091] Slave database: Deploy 2-3 slave databases to synchronize index data with the master database and handle query requests; improve query throughput through read-write separation.

[0092] Database sharding and table partitioning strategy: shard databases by hash modulo of cluster ID, and partition tables by cell ID range; for example: shard the cluster ID hash value %3 to 3 master databases, and split the data tables in each master database according to the range of the first 6 digits of the cell ID (e.g., 000000-100000 is a table).

[0093] See Figure 2 This application also provides a user information query device, the user information query device comprising: A raw crawling data acquisition module, which is used to acquire raw crawling data; A standardized identity cell set acquisition module is used to generate a standardized identity cell set based on the original crawled data. A convergent cell cluster set acquisition module is used to perform cell group cooperative evolution on a standardized identity cell set to obtain a convergent cell cluster set. The dynamic association index database acquisition module is used to dynamically index the evolution results of the convergent cell cluster set, thereby acquiring the dynamic association index database. User request information acquisition module, the user request information acquisition module is used to acquire user request information; The query result acquisition module is used to perform a query from the dynamic association index database based on the user request information, thereby obtaining the query result.

[0094] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for querying user information, characterized in that, The user information query method includes: Obtain the raw crawled data; Generate a standardized set of identity cells based on the raw crawled data; The standardized identity cell set is subjected to cell group co-evolution to obtain a convergent cell cluster set; Dynamic index mapping of evolution results is performed on the convergent cell cluster set to obtain a dynamic association index database; Obtain user request information; The system retrieves query results by querying the dynamically associated index database based on the user's request information.

2. The user information query method as described in claim 1, characterized in that, The process of generating a standardized identity cell set based on the original crawled data includes: The original crawled data is parsed using multi-source heterogeneous data parsing to obtain the original identity unit set. The original identity unit set includes multiple cells, and each cell includes a unique identifier ID, data content, data type, and original source address. The initial identity cell set with attributes is generated by dynamically assigning contextualized cell attributes to each cell in the original identity unit set. The initial identity cell set with attributes includes multiple cells, each containing a unique ID, basic attributes, feature attributes, and associated attributes. The initial set of identity cells with attributes is standardized by cell standardization and core candidate screening to obtain a standardized set of identity cells.

3. The user information query method as described in claim 2, characterized in that, The original crawled data includes text-based crawled data, visual-based crawled data, and structured crawled data; The step of parsing the original crawled data from multiple sources to obtain the original set of identity units includes: The text-based crawled data is parsed to obtain the original identity units of the text. The visual data crawled is parsed to obtain the original visual identity units; The structured crawled data is parsed to obtain the original identity units of the structured class; wherein the original identity units of the text class, the original identity units of the visual class, and the original identity units of the structured class constitute the set of original identity units.

4. The user information query method as described in claim 3, characterized in that, The step of dynamically assigning scenario-based cell attributes to each cell in the original identity unit set to generate an initial identity cell set with attributes includes: Each cell is labeled with contextualized basic information, which includes textual context, visual context, and structured context. When the contextualized basic information of a cell is a text scene, the format compliance feature and semantic coherence feature of the cell are extracted, and these format compliance feature and semantic coherence feature are the feature attributes of the cell. When the contextualized basic information of a cell is a visual scene, the text content and pixel distribution features of the cell are extracted, and the text content and pixel distribution features are fused into a text-visual composite feature, which is used as the feature attribute of the cell. When the contextualized basic information of a cell is a structured scene, the field value features of the cell are extracted, and these field value features are used as the feature attributes of the cell. The text scene association potential value of the cell is generated based on the format compliance feature and semantic coherence feature, and the text scene association potential value is used as the association attribute of the cell. The visual scene association potential value of the cell of the visual scene is generated based on the visual composite features of the text, and the visual scene association potential value is used as the association attribute of the cell. Based on the field value characteristics, a structured scene association potential value is generated for the cell of the structured scene, and this structured scene association potential value is used as the association attribute of the cell.

5. The user information query method as described in claim 4, characterized in that, The text scene association potential value, visual scene association potential value, and structured scene association potential value are all calculated using the following formula: ； in, For format compliance; For field co-entropy; The scene-adaptive entropy; t is the time decay factor; For the number of field collaborations.

6. The user information query method as described in claim 5, characterized in that, The step of performing cell-group cooperative evolution on the standardized identity cell set to obtain a convergent cell cluster set includes: For each cell in the standardized identity cell set, evolution rules and cell attributes are bound to the cell in a contextualized manner to obtain a contextualized binding rule set. Each cell in the contextualized binding rule set includes a field type, sensitivity label, exclusive evolution rule parameters, and a list of field association weights with potential associated cells. The evolutionary environment is initialized and the neighborhood is defined based on the scenario-based binding rule set, thereby generating a three-dimensional evolutionary space model and an initial association weight matrix; Based on the three-dimensional evolution space model and the initial correlation weight matrix, multiple rounds of iterative evolution are performed to obtain the final convergent cell cluster set.

7. The user information query method as described in claim 6, characterized in that, The process involves binding each cell in the standardized identity cell set with evolution rules and cell attributes in a contextualized manner to obtain a set of contextualized binding rules, including: Iterate through each cell in the standardized identity cell set and extract the identity category information of each cell; Based on the identity information category of each cell, a sensitivity label is generated for each cell. Based on the sensitivity labels, set exclusive evolutionary rules for each cell; Generate field association weights for every two arbitrary cells; Each cell in the scenario-based binding rule set includes a field type, sensitivity label, exclusive evolution rule parameters, and a list of field association weights with potential related cells.

8. The user information query method as described in claim 7, characterized in that, The process of initializing the evolutionary environment and defining the neighborhood based on a scenario-based binding rule set, thereby generating a three-dimensional evolutionary space model and an initial association weight matrix, includes: Each cell in the scenario-based binding rule set is mapped to a cell space, and all cells are mapped to a three-dimensional evolution space according to their source carrier, thereby obtaining a three-dimensional evolution space model. Based on the three-dimensional evolution space model, cells that satisfy the conditions of spatial distance ≤ neighborhood radius and field association weight ω ≥ 0.4 are selected as neighborhood cell pairs; Calculate the initial association weight value for each pair of neighboring cells to form the initial association weight matrix.

9. The user information query method as described in claim 8, characterized in that, The process of obtaining the final convergent cell cluster set by performing multiple rounds of iterative evolution based on the three-dimensional evolution space model and the initial correlation weight matrix includes: Based on the three-dimensional evolution space model and the initial correlation weight matrix, the following first-stage iterative operation is performed until the first iteration requirement is met: Calculate the overall similarity between each pair of neighboring cells; Update the association weight matrix of each neighborhood cell pair whose comprehensive similarity satisfies the first similarity preset condition; After updating the correlation weight matrix, each cell is subjected to a double validity screening to obtain noisy cells and temporary cell clusters. After the first stage of iteration is completed, the temporary cell clusters obtained after the first stage of iteration are subjected to cross-page cell co-evolution to obtain the final convergent cell cluster set.

10. A user information query device, characterized in that, The user information query device includes: A raw crawling data acquisition module, which is used to acquire raw crawling data; A standardized identity cell set acquisition module is used to generate a standardized identity cell set based on the original crawled data. A convergent cell cluster set acquisition module is used to perform cell group cooperative evolution on a standardized identity cell set to obtain a convergent cell cluster set. The dynamic association index database acquisition module is used to dynamically index the evolution results of the convergent cell cluster set, thereby acquiring the dynamic association index database. User request information acquisition module, the user request information acquisition module is used to acquire user request information; The query result acquisition module is used to perform a query from the dynamic association index database based on the user request information, thereby obtaining the query result.

Images